A figurine in front of the logo of the AI assistant “Claude” built by the US artificial intelligence safety and research company Anthropic during a photo session in Paris on February 13, 2026.
Joel Saget/AFP via Getty Images
A federal judge signed off Monday on Anthropic’s $1.5 billion settlement with a class of authors and publishers who sued the AI company for downloading millions of pirated books to train its Claude chatbot — delivering the largest copyright class-action payout in American legal history and, for the first time, a market-set price for wholesale AI ingestion of copyrighted works.
But the most consequential thing the ruling settled may be what it did not settle. The fair-use ruling that preceded it — issued by then-presiding Judge William Alsup last June — was a single district-court decision. Because Anthropic chose to settle rather than let the case proceed to an appeals court, that ruling will never become binding precedent. The dozens of AI copyright cases now moving through federal courts, from a new publisher lawsuit against Google filed just last week to ongoing battles involving OpenAI, Meta, and Midjourney, remain legally free to reach entirely different conclusions.
That distinction — between a case being resolved and a legal question being answered — defines what Monday’s ruling actually accomplished.
Judge Approves $1.5B Payout, Rejects Author Objections
U.S. District Judge Araceli Martínez-Olguín, who inherited the case after Alsup’s retirement, signed the final approval order in San Francisco, affirming that the settlement was “fair and adequate” and overruling objections from a minority of authors who argued the payout undervalued their claims. The judge found that complaints about the settlement’s size were “not grounded in a realistic assessment of the overall risks and rewards of a trial,” according to the court’s final approval order.
The math behind the settlement: roughly $3,000 for each of approximately 500,000 works covered by the class, paid out to the authors and publishers who hold rights to those titles. More than 92% of eligible copyright holders submitted claims. As a condition of the settlement, Anthropic agreed to destroy the pirated files it had accumulated — a significant operational concession for a company that had assembled one of the largest private book corpora in existence.
Anthropic Deputy General Counsel Aparna Sridhar welcomed the conclusion: “We are pleased that more than 91% of authors and publishers covered by the settlement have claimed their share of the payment, and we’re looking forward to bringing this matter to a close.” Plaintiffs’ lead attorney Justin Nelson called it a milestone: “It is the largest known copyright recovery in history. We look forward to making distributions to the Class as promptly as possible.”
What Anthropic Actually Did: Books Bought, Stripped, and Pirated
The case, Bartz et al. v. Anthropic PBC (Case No. 3:24-cv-05417, N.D. Cal.), was filed in 2024 by named plaintiffs including novelists Andrea Bartz, Charles Graeber, and Kirk Wallace Johnson. The core allegation was that Anthropic, backed by investors Amazon and Alphabet, had assembled its book training corpus through two methods: one legal, one not.
The legal method: Anthropic purchased millions of print copies in bulk, stripped the books’ bindings, cut the pages, and scanned them into machine-readable digital form. The illegal method: the company also downloaded hundreds of thousands of digital books from shadow libraries — specifically Library Genesis and Pirate Library Mirror, online repositories that host pirated copies of books and academic papers in deliberate defiance of copyright law.
In June 2025, Judge Alsup issued what amounted to a split verdict. He ruled that training an AI model on copyrighted books constituted transformative fair use under 17 U.S.C. § 107 — a ruling widely cited as the AI industry’s first major fair-use victory. But he held separately that maintaining a permanent “central library” of more than seven million pirated books fell outside that protection. The piracy was the act that exposed Anthropic to massive statutory liability: under 17 U.S.C. § 504(c)(2), willful copyright infringement carries statutory damages of up to $150,000 per work. With seven million-plus books potentially in scope, the theoretical damages ceiling stretched into the hundreds of billions of dollars — a number that concentrated minds on both sides of the negotiating table.
A trial to determine how much Anthropic owed was scheduled for December 2025. Anthropic settled before that trial began.
Why Other AI Companies Are Not Off Court
The settlement’s most important limitation — for every other AI company, every copyright holder, and every federal judge now presiding over a related case — is structural: a class-action settlement does not create precedent.
Alsup’s fair-use ruling was a trial-court decision from the Northern District of California. Because Anthropic settled rather than appeal, that ruling never reached the Ninth Circuit or the Supreme Court. Other federal courts are not bound by it. The Southern District of New York, which is now hearing a new class-action filed by Hachette Book Group, Cengage Learning, Elsevier, and author Scott Turow against Google over alleged Gemini training-data piracy (filed July 10, 2026), is entirely free to reach the opposite conclusion on whether AI training constitutes transformative fair use.
That matters because the Hachette-Google complaint contains details the Bartz case did not: Google allegedly removed or altered copyright information on ingested works to “conceal that its Gemini Models were trained on stolen materials” — a claim that goes beyond careless data sourcing toward alleged intentional concealment. An internal Google document cited in the filing reportedly described the practice as “highly problematic” and warned of potential fines between $10 billion and $100 billion.
Also still active: Kadrey et al. v. Meta Platforms (N.D. Cal.), where a court granted a partial motion to dismiss on fair use for the LLM-training claim while piracy-related claims — specifically allegations involving torrented books — remain live. The pattern is becoming a template: training use may be defensible; sourcing from shadow libraries is not.
Legal experts have flagged the absence of binding appellate law as the defining uncertainty for every remaining case. One legal expert noted that large AI companies “need to be concerned with the potential legal risk — the intellectual property-based risks of these suits — but they also need to be concerned with the potential tarnishment of their brands when they are alleged to have pirated, stolen, or used without permission or compensation others’ creative works.” The U.S. Copyright Office noted in a May 2025 report that it was skeptical all AI training constitutes transformative fair use, particularly when done commercially and when model outputs compete directly with the original licensed works.
What Does $3,000 Per Book Actually Mean for the Creative Economy?
For the broader creative industry, the settlement introduces something that had been conspicuously absent from the AI training-data debate: a market price. At roughly $3,000 per work across 500,000 titles, the settlement represents the first judicially endorsed per-unit valuation for AI training use of copyrighted books.
That number reads very differently depending on who is doing the reading.
Critics within the author community have argued that $3,000 is a small fraction of what a commercially successful book earns over its lifetime, and that the settlement effectively prices a market rate for wholesale ingestion by well-capitalized AI labs — a rate most authors would not have agreed to if offered it directly as a license fee. For Anthropic, whose valuation has grown dramatically since the settlement was announced in September 2025 — reaching approximately $1.2 trillion on private secondary markets by July 2026 — the $1.5 billion total represents less than one-tenth of 1% of its estimated worth.
Supporters frame it differently. Cecilia Ziniti, a tech industry lawyer and former Ninth Circuit clerk, described the outcome as the start of a structural realignment: “This settlement marks the beginning of a necessary evolution toward a legitimate, market-based licensing scheme for training data. It’s not the end of AI, but the start of a more mature, sustainable ecosystem where creators are compensated, much like how the music industry adapted to digital distribution.”
The music analogy is instructive: Napster was shut down by litigation in 2001, and the streaming licensing frameworks that replaced it took more than a decade to mature. The Anthropic settlement sets one data point. Whether it becomes the floor for that emerging market — or an outlier that other AI companies try to avoid matching — depends largely on how the Google, Meta, and OpenAI cases develop.
Nick Schneider, an attorney at Eckert Seamans, noted that the settlement sends a clear message to the rest of the industry: “newer AI companies must invest in legitimate data acquisition, negotiate licensing agreements, and carefully document all training materials.” The key legal line drawn: AI companies cannot use shadow-library piracy to build their training corpora and then invoke training-use fair use to retroactively sanitize the acquisition. The two acts are legally separable — and only one of them gets fair-use protection.
Objectors, Opt-Outs, and What Comes Next for Anthropic
The settlement was not unanimous within the plaintiff class. Some authors and publishers objected that it undervalued their claims, overcompensated the attorneys, or wrongly excluded certain copyright holders — all arguments Judge Martínez-Olguín rejected. She awarded the plaintiffs’ attorneys approximately $122 million in fees and litigation costs from the total fund.
Authors who opted out of the settlement filed separate lawsuits against Anthropic that remain pending. Additionally, Anthropic faces a separate $3 billion music-copyright lawsuit filed in January 2026 by Universal Music Group, Concord Music Group, and ABKCO — naming both the company and CEO Dario Amodei and co-founder Benjamin Mann individually — over alleged use of song lyrics and sheet music to train Claude.
Anthropic, which did not admit liability, emphasized the aspect of the ruling that it preferred: the finding that AI training on books is fair use. “We reached this settlement in 2025, after the court’s landmark ruling that training AI on books is fair use under copyright law — which remains the law today,” Sridhar said. What Anthropic did not address publicly is the limitation of that claim: “remains the law today” in the Northern District of California, in a trial-court ruling that cannot bind any other court.
The $1.5 billion figure will reverberate through every deposition room and boardroom that touches AI and intellectual property for years to come. What it will not do is answer the question that the industry most needs answered: whether courts will ultimately accept or reject the proposition that training a commercial AI model on copyrighted works is transformative fair use. That question remains open, and the legal terrain where it will be resolved stretches well beyond San Francisco.
Frequently Asked QuestionsDoes this settlement mean AI companies can legally use copyrighted books to train their models?
Not as a binding legal precedent. The settlement was preceded by a ruling from Judge Alsup that training an AI on copyrighted books may qualify as transformative fair use — but Alsup’s decision was a single district-court ruling in California, and the settlement means it will never be reviewed by an appeals court. Other federal courts, including the New York court now hearing a lawsuit against Google over Gemini training data, are free to reach a different conclusion. The U.S. Copyright Office has also expressed skepticism that all commercial AI training qualifies as fair use, particularly when AI outputs compete with original licensed works.
How much will individual authors and publishers actually receive from the settlement?
The settlement distributes approximately $3,000 per covered work across an estimated 500,000 titles. More than 92% of eligible copyright holders filed claims. The specific per-claimant amount depends on how many works a rights holder owns in the covered corpus. Authors who opted out of the class are pursuing separate lawsuits and will not receive payments from the settlement fund.
What specifically did Anthropic do that made it liable — wasn’t AI training covered by fair use?
The ruling distinguished between two separate acts. Training an AI on copyrighted text may be fair use under Alsup’s analysis — that portion of the claim was decided in Anthropic’s favor. What was not protected: downloading and permanently storing more than seven million books from shadow libraries like Library Genesis and Pirate Library Mirror, which constitute copyright piracy regardless of how the books were subsequently used. Piracy is not laundered clean by a fair-use defense on the downstream training use. It was the sourcing method — not the training itself — that created Anthropic’s liability and exposure to statutory damages of up to $150,000 per work.
What does the Anthropic settlement mean for the other AI copyright cases still in court?
It sets a powerful financial benchmark and a clear warning about data-sourcing from shadow libraries, but it does not resolve the legal question for anyone else. The Google, Meta, OpenAI, and Midjourney cases remain pending, each governed by its own facts and, in Google’s case, a different federal court in New York that is not bound by California precedent. The Anthropic settlement may accelerate settlement discussions in those cases — but it may also set a per-work price that some defendants will fight rather than accept.